fix(openai): retain cache-write usage telemetry

[zzwong]

Summary

Refs #621.

Retain the Responses API's usage.input_tokens_details.cache_write_tokens in the existing nullable CacheCreate usage field. Omitted/null counts remain unavailable; an explicit zero remains a reported zero.

• Preserve OpenAI's inclusive input total and independently reported cache reads.
• Add nine raw/nullable telemetry cases and six malformed-type/overflow cases.
• Keep raw negative and overlapping counts unchanged. They are telemetry, not validated billable usage.
• Verify the bundled OpenAI context-banded estimator remains unavailable and Anthropic's separate cache buckets retain their existing behavior.

The field and inclusive input semantics are documented in OpenAI's prompt-caching guide.

Remaining cost work

This is a telemetry-only slice of #621. It adds no new estimator activation and does not complete OpenAI or Codex cost estimation.

Complete request histories are still needed across validation and transport retries. Context bands and observed service tiers must be selected per request before aggregation; the catalog currently has no numeric context boundary. Ordinary OpenAI input must exclude both cache reads and cache writes before applying rates. Missing, inconsistent, or otherwise incomplete billing evidence must leave future estimates unavailable.

No lifecycle, durable-schema, catalog, pricing-formula, or Codex changes are included.

Focused local validation

The exact proposed source passed bounded offline checks with keyring_nopassage and an isolated cold build cache:

• Parser/characterization normal: 5 top-level tests, 28 passing test/subtest results.
• Selected direct-API regression normal: 16 top-level tests, 52 passing test/subtest results.
• Existing pricing package: 24 passing test/subtest results.
• The same selected direct-API scope with -race: 16 top-level tests, 52 passing test/subtest results, no race diagnostics.

All nine telemetry cases, six malformed/overflow cases, three bundled-price guards, and the Anthropic bucket regression were observed. No selected tests failed or skipped. Formatting and patch-application checks passed; source stayed unchanged throughout execution.

Full adapter-package tests, repository-wide tests, lint, standalone build, live-provider calls, and a baseline-red execution were not run locally. These focused results do not establish merge readiness.

Source scope

Based on 999e59ddfc74a67ffeef3c4b4a78a8528365afe6. The exact tested Git tree is 1701116532002ffa0c8e3ab1653728d9bf1f15b0; only internal/llmadapters/api.go and internal/llmadapters/api_test.go differ. This remains a draft, and #621 remains open.

[monit-reviewer]

Automated PR Review

Reviewed commit: 29edd5c6877b
Profile: open-cli-collective-review - Posting as: monit-reviewer

Summary

Reviewer	Findings
go:implementation-tests	0

Reviewer Coverage

Complete coverage details are retained in the local coverage.json artifact.

• go:implementation-tests — complete (broad); scope: 2 files; examples: internal/llmadapters/api.go, internal/llmadapters/api_test.go; body-inspected: 2 files; examples: internal/llmadapters/api.go, internal/llmadapters/api_test.go; skipped: none; context (not coverage): 1 file; examples: internal/pricing/pricing.go; constraints: Static review only; tests were inspected but not executed.

0 PR discussion threads considered. 0 summarized; 0 resolved.

---

Completed in 27s | $0.03 | openai-codex/gpt-6.1-sol | cr 0.10.328

Field	Value
Model	openai-codex/gpt-6.1-sol
Reviewers	go:implementation-tests
Engine	pi_rpc · openai-codex/gpt-6.1-sol
Reviewed by	cr · monit-reviewer
Duration	27s wall · 17s compute
Cost	$0.03
Tokens	10.2k in / 268 out

Per-workstream usage

• orchestrator-selection — openai-codex/gpt-6.1-sol
  • In: 5.6k
  • Out: 121
  • Cache read: 0
  • Cache create: 0
  • Cost: $0.01
  • Duration: 3s
• go:implementation-tests — openai-codex/gpt-6.1-sol
  • In: 977
  • Out: 100
  • Cache read: 19.6k
  • Cache create: 0
  • Cost: $0.00
  • Duration: 10s
• orchestrator-rollup — openai-codex/gpt-6.1-sol
  • In: 3.6k
  • Out: 47
  • Cache read: 0
  • Cache create: 0
  • Cost: $0.01
  • Duration: 2s

[monit-reviewer]

Automated PR Review

Reviewed commit: 884955e14282
Profile: open-cli-collective-review - Posting as: monit-reviewer

Summary

Reviewer	Findings
go:implementation-tests	0

Reviewer Coverage

Complete coverage details are retained in the local coverage.json artifact.

• go:implementation-tests — complete (constrained); scope: 2 files; examples: internal/llmadapters/api.go, internal/llmadapters/api_test.go; body-inspected: 2 files; examples: internal/llmadapters/api.go, internal/llmadapters/api_test.go; skipped: none; context (not coverage): 2 files; examples: docs/development.md, internal/pricing/pricing.go; constraints: Static inspection only; tests and build commands were not executed.

0 PR discussion threads considered. 0 summarized; 0 resolved.

---

Completed in 25s | $0.01 | openai-codex/gpt-6.1-sol | cr 0.10.328

Field	Value
Model	openai-codex/gpt-6.1-sol
Reviewers	go:implementation-tests
Engine	pi_rpc · openai-codex/gpt-6.1-sol
Reviewed by	cr · monit-reviewer
Duration	25s wall · 14s compute
Cost	$0.01
Tokens	4.6k in / 130 out

Per-workstream usage

• go:implementation-tests — openai-codex/gpt-6.1-sol
  • In: 992
  • Out: 80
  • Cache read: 22.9k
  • Cache create: 0
  • Cost: $0.01
  • Duration: 11s
• orchestrator-rollup — openai-codex/gpt-6.1-sol
  • In: 3.6k
  • Out: 50
  • Cache read: 0
  • Cache create: 0
  • Cost: $0.01
  • Duration: 3s